Manipulability-Augmented Next-Best-Configuration Exploration Planner for High-DoF Manipulators

This letter presents MA-NBCP, a novel hierarchical framework targeting autonomous exploration and inspection for high-DoF manipulators. MA-NBCP iteratively selects the manipulability-augmented next-best-configuration for exploring unknown regions surrounding a manipulator, while providing collision-avoidance guarantee. Toward developing MA-NBCP, an efficient exploration information system (EIS) is first built that dynamically maintains critical, extensible information to facilitate the exploration planning process. Leveraging EIS, the higher level of MA-NBCP selects the best exploration subregion based on an angle-weighted metric. At the lower level, a Fibonacci grids-based spherical uniform sampling strategy generates many candidate viewpoints. The process yields a diverse set of sensor-robot configurations, which are subsequently ranked based on the information gain of the viewpoint and manipulability index of the corresponding robot configuration to jointly determine the next-best-configuration. To further speed up run-time lookup, a database containing high-manipulability robot configurations is pre-built and integrated into EIS. As a result, MA-NBCP can efficiently carry out autonomous collision-free exploration of unknown environments. Thorough simulation and real hardware (over a 7-DoF manipulator equipped with a depth camera) in highly confined settings demonstrate that MA-NBCP has significant advantages over the current SOTA approaches in terms of exploration time and distance travelled in the joint space (specifically, 56% and 63% better on average, respectively), as well as the mean manipulability index of intermediate configurations at exploration iterations (79% higher on average).

I. Introduction

With vast potentials toward real-world applications, e.g., human-robot collaboration [1], medical robotics [2], assisted living, and so on, robotic manipulators integrated with vision systems have attracted significant research attention. For effective task execution in rapidly evolving environments, e.g., kitchens, hospital storage rooms, and flexible manufacturing workcells, a manipulator must be able to quickly understand its immediate surrounding environment, which may change on a daily basis even if the robot itself remains stationary. At the hardware level, this is best achieved by coupling vision to the end-effector of the manipulator, avoiding the inflexibility of fixed cameras, which have limited field-of-view (FoV) and their views may also be blocked by the robot and changing obstacles in the environment. Our study focuses on developing exploration algorithms for enabling high-DoF eye-on-hand manipulation systems to rapidly and accurately map out their immediate surroundings, ensuring their safe on-demand deployment, as shown in Fig. 1.